Search

Explainability

The ability to describe mechanisms behind an AI system’s operation. A useful explanation depends on its audience and purpose; an explanation of a score is not automatically an accurate, legally sufficient adverse-action notice.

False positive

A positive model or rule signal when the target condition is absent—for example, flagging a legitimate transaction as fraud. The false-positive rate uses actual negatives as its denominator; the share of alerts that are wrong is a different metric.

Definition

…when the target condition is absent—for example, flagging a legitimate transaction as fraud. The false-positive rate uses actual negatives as its denominator; the share of alerts that are wrong is a different metric.

Model drift

Changes over time in a model’s inputs, the relationship between inputs and outcomes, or observed performance. Data drift describes a changing input population; it may warrant investigation even before performance deterioration is measurable.

Model validation

Evaluation of whether a model is conceptually sound, implemented appropriately and performing adequately for its intended use. Testing and challenge should reflect the model’s risks; validation does not guarantee future accuracy or eliminate the need for monitoring.

Synthetic identity

A fabricated person or entity assembled from combinations of identifying information. Synthetic identity fraud uses that identity for dishonest gain; it differs from simply taking over an existing real person’s account.

Definition

…or entity assembled from combinations of identifying information. Synthetic identity fraud uses that identity for dishonest gain; it differs from simply taking over an existing real person’s account.